Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING 1
Cognitive-LPWAN: Towards Intelligent WirelessServices in Hybrid Low Power Wide Area Networks
Min Chen , Senior Member, IEEE, Yiming Miao , Xin Jian , Xiaofei Wang , Senior Member, IEEE,
and Iztok Humar , Senior Member, IEEE
Abstract—The relentless development of the Internet of Things1
(IoT) communication technologies and the gradual maturity of2
artificial intelligence (AI) have led to a powerful cognitive com-3
puting ability. Users can now access efficient and convenient4
smart services in smart-city, green-IoT, and heterogeneous net-5
works. AI has been applied in various areas, including the intel-6
ligent household, advanced health-care, automatic driving, and7
emotional interactions. This paper focuses on current wireless-8
communication technologies, including cellular-communication9
technologies (4G, 5G), low-power wide-area (LPWA) technolo-10
gies with an unlicensed spectrum (LoRa, SigFox), and other11
LPWA technologies supported by 3GPP working with an autho-12
rized spectrum (EC-GSM, LTE-M, NB-IoT). We put forward13
a cognitive LPWA-network (Cognitive-LPWAN) architecture to14
safeguard stable and efficient communications in a heteroge-15
neous IoT. To ensure that the user can employ the AI efficiently16
and conveniently, we realize a variety of LPWA technologies17
to safeguard the network layer. In addition, to balance the18
demand for heterogeneous IoT devices with the communication19
delay and energy consumption, we put forward the AI-enabled20
LPWA hybrid method, starting from the perspective of traffic21
control. The AI algorithm provides the smart control of wireless-22
communication technology, intelligent applications and services23
for the choice of different wireless-communication technologies.24
As an example, we consider the AIWAC emotion interaction25
system, build the Cognitive-LPWAN and test the proposed AI-26
enabled LPWA hybrid method. The experimental results show27
that our scheme can meet the demands of communication-delay28
Manuscript received August 13, 2018; revised September 8, 2018; acceptedSeptember 23, 2018. This work was supported in part by the NaturalScience Foundation of China under Grant 61501065, Grant 61702364, GrantU1705261, and Grant 61572220, in part by the National Key Researchand Development Program of China under Grant 2018YFC1314600 andGrant 2018YFC0809803, in part by the Fundamental Research Funds for theCentral Universities under Grant HUST: 2018KFYXKJC045, in part by theSlovenian Research Agency (research core funding No. P2-0246), and in partby Director Fund of WNLO. The associate editor coordinating the review ofthis paper and approving it for publication was H. Ji. (Corresponding author:Xiaofei Wang.)
M. Chen is with the Department of Computer Science and Technology,Huazhong University of Science and Technology, Wuhan 430074, China, andalso with the Wuhan National Laboratory for Optoelectronics, Wuhan 430074,China (e-mail: [email protected]).
Y. Miao is with the School of Computer Science and Technology,Huazhong University of Science and Technology, Wuhan 430074, China(e-mail: [email protected]).
X. Jian is with the College of Communication Engineering, ChongqingUniversity, Chongqing 400030, China (e-mail: [email protected]).
X. Wang is with the Tianjin Key Laboratory of Advanced Networking,College of Intelligence and Computing, Tianjin University, Tianjin 300350,China (e-mail: [email protected]).
I. Humar is with the Faculty of Electrical Engineering, University ofLjubljana, 1000 Ljubljana, Slovenia (e-mail: [email protected]).
Digital Object Identifier 10.1109/TGCN.2018.2873783
applications. Cognitive-LPWAN selects appropriate communica- 29
tion technologies to achieve a better interaction experience. 30
Index Terms—Artificial intelligence, low-power wide-area net- 31
work, LoRa, LTE, NB-IoT. 32
I. INTRODUCTION 33
THE INTERNET of Things (IoT) forms connections 34
between people and things, as well as between things 35
and things based on wired or wireless communication tech- 36
nologies. The IoT establishes a thing-thing Internet with a 37
wide geographical distribution and provides innovative appli- 38
cations and services [1]. At present, global telecommunication 39
operators have established mobile cellular networks with wide 40
coverage [2]. Although the 2G, 3G, 4G and other mobile 41
cellular technologies [3] support wide coverage and high 42
transmission rates, they suffer from various disadvantages, 43
such as large power consumption and high costs. According 44
to Huawei’s analysis report released in February 2016, the 45
number of connections between the things and things on a 46
global mobile cellular network occupies only 10% of all the 47
connections [4]. 48
The crucial design purpose of mobile cellular- 49
communications technology is to improve the interpersonal 50
communication efficiency, since the current capacity of 51
the mobile cellular network is not sufficient to support 52
the massive connections between things and things. Up to 53
2025 the number of connections with industrial wireless 54
sensing, tracking and control devices will reach to 500 55
million [1]. Therefore, the IoT should be able to provide 56
end-users with convenient and efficient intelligent services 57
and open access to historical data, while integrating a large 58
number of heterogeneous terminal devices (the terminals of 59
things) in a transparent and seamless manner. With such 60
services, artificial intelligence (AI) systems can monitor 61
the users and their surroundings more efficiently [5]. This 62
results in smart cities, smart homes, autopilot system [6] and 63
health-monitoring applications, and yields a greener, more 64
environmentally friendly, and more efficient IoT ecosystem 65
for more cost-effective systems. 66
The diversity in terms of requirements and technologies for 67
the Internet of Things has given rise to the heterogeneity of the 68
network structures and the instability of the design solutions. 69
Although traditional cellular networks provide long-distance 70
coverage, they hardly provide energy-efficient (EE) connec- 71
tivity due to their complex modulation and multiple access 72
2473-2400 c© 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2 IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING
Fig. 1. Historical evolution of wireless-communication technologies.
technologies [7]. As a result of the continuous development73
of the IoT its communication technologies have become more74
mature. These technologies can be divided into two main cat-75
egories according to their transmission distances, as follows.76
The first category includes short-range communications tech-77
nologies with representative technologies of ZigBee, WiFi,78
Bluetooth, Z-wave, etc. and typical application scenarios such79
as smart homes. The second category is the wide-area network80
communication technology defined as a low-power wide-area81
network (LPWAN), with typical application scenarios such as82
autonomous driving. The wide-area-network communication83
technology is required in low-speed businesses as a revolu-84
tionary IoT access technology, where typical wireless local85
area networks (WLANs), i.e., the WiFi, are less costly but86
suffer from a limited coverage distance. Therefore, LPWA87
technology is very promising thanks to its long-distance88
coverage, low power consumption, low data rates and low89
costs [8].90
LPWA technologies can be classified into two categories.91
The first category includes the LoRa, SigFox and other92
technologies that work with unlicensed spectra. The second93
category includes 2/3/4G cellular communication technologies94
with licensed spectra based on the 3GPP, such as EC-GSM,95
LTE-M, NB-IoT, etc. According to such diverse wireless-96
communication technologies, here we propose the cognitive97
low-power wide-area network (Cognitive-LPWAN). Such a98
network aims at smart cities, green IoT and other hetero-99
geneous networks in AI applications such as smart home,100
health monitoring, automatic driving and emotional interac-101
tion. We show that the Cognitive-LPWAN provides users with102
more efficient and convenient intelligent services by taking103
advantage of LPWA technologies.104
Generally speaking, the purpose of a LPWAN is to pro-105
vide long-distance communication, i.e., about 30 kilometers106
coverage in rural areas and 5 kilometers coverage in urban107
areas. Meanwhile, many IoT devices are characterized by a108
service time of more than 10 years. Therefore, we need to109
improve the transmission range and power consumption of110
the LPWAN in order to adapt it to IoT applications that are111
highly extensible, e.g., the intelligent monitoring infrastruc-112
ture, in which only a small portion of the data is transmitted.113
Thus, two possible technologies are proposed to solve these114
two problems. The first one is an ultra-narrowband technology115
that enhances the signal-to-noise ratio (SNR) by focalizing116
the signals in a narrow band. The narrowband IoT (NB-117
IoT) [9] is an implementation example of this approach.118
Another approach is to use a coding gain to alleviate the119
high noise and power consumption in wideband receivers,120
such as the Long-Range Wireless Communication (LoRa) [10] 121
technology that increases the transmission distance by increas- 122
ing the power consumption. However, wireless-communication 123
technologies with unlicensed spectra can conflict with other 124
business flows with respect to channel collisions and spectrum 125
occupation if they are not well controlled [11]. If this technol- 126
ogy is abandoned for these reasons, a market with hundreds 127
of millions of IoT terminals will also be lost. 128
Artificial intelligence technologies have become more 129
mature, and we have now more powerful cognitive comput- 130
ing capabilities in regards to business perception at the user 131
level, intelligent transmission at the network level and big- 132
data analysis in the cloud. Aiming at traffic control, this paper 133
proposes a new, intelligent solution for the Cognitive-LPWAN 134
architecture, i.e., the AI-enabled LPWA hybrid method. We 135
use the AI algorithm in a data and resource cognitive engine. 136
LPWA technology has been widely applied to green IoT, 137
so hopefully the new LPWAN architecture and its solutions 138
for wireless-technology selection will contribute to the IoT 139
ecosystem. 140
The main contributions of this paper are as follows: 141
• We investigate several typical LPWA technologies, 142
including the LoRa, SigFox and other wireless- 143
communication technologies with unlicensed spectra, and 144
3GPP-supported 2/3/4G cellular-communication tech- 145
nologies with licensed spectra such as EC-GSM, LTE 146
Cat-m, NB-IoT, etc. 147
• We propose the Cognitive-LPWAN as a combination 148
of multiple LPWA technologies, ensuring more effi- 149
cient and convenient user experiences in the AI services 150
on the network layer. As a result, stable and efficient 151
communication is guaranteed between the people and 152
things, people and people, and things and things in the 153
heterogeneous IoT. 154
• We puts forward the AI-enabled LPWA hybrid method, 155
starting off from the perspective of the flow control, 156
which makes a balance between the communication time 157
delay and energy consumption in heterogeneous IoT 158
devices. By using the AI algorithm, we achieve smart 159
control for communication traffic, intelligent applications 160
and services for the choice of the wireless-communication 161
technology. 162
• We establish an experimental platform according to the 163
AIWAC emotion interaction system, and compare the AI- 164
enabled LPWA hybrid method with the single-technology 165
mode. The experimental results show that our method can 166
choose the communication technology accordingly and 167
demonstrates proper transmission-delay performance. 168
CHEN et al.: COGNITIVE-LPWAN: TOWARDS INTELLIGENT WIRELESS SERVICES IN HYBRID LOW POWER WIDE AREA NETWORKS 3
The rest of this paper is organized as follows. Section II169
summarizes the existing heterogeneous, low-power,170
wide-area-network technologies. Section III introduces171
Cognitive-LPWAN and proposes the solution of multiple172
wireless-communication technologies in the smart-cities173
environment. Next, we present the AI-enabled LPWA hybrid174
method modeling in Section IV. Section V demonstrates the175
building testbed through different wireless-communication176
technologies based on baby robots. This building testbed is177
used to test the performance of the proposed AI-enabled178
LPWA hybrid method. We then discuss open issues for the179
future in Section VI. Finally, Section VII summarizes the180
paper.181
II. HETEROGENEOUS LOW-POWER182
WIDE-AREA-NETWORK TECHNOLOGY183
A. SigFox184
SigFox is provided by the SIGFOX company that was185
founded by the Ludovic Le Moan and Christophe Fourtet186
in 2009. As a LPWA technology worked on non-licensed187
spectrum, SigFox has been rapidly commercialized and pro-188
vides network devices with ultra-narrowband technology [12].189
Different from mobile communication technology, it is a pro-190
tocol that aims to create wireless IoT special networks with191
low power consumption and low costs. More concretely, the192
maximum length of each message on a device that operates193
based on the Sigfox network is nearly 12 bytes due to the 50194
microwatts upper limit of one-way communications. And no195
more than 150 messages will be sent per day by each device.196
Moreover, the coverage provided by a SigFox network can197
reach up to 13 kilometers.198
B. LoRa199
LoRa is currently one of the most common LPWA technolo-200
gies worked on non-licensed spectrum, which is provided by201
SemTech [10]. The main characteristics of LoRa wireless tech-202
nology include 20 km coverage range, 10,000 or even millions203
of nodes, 10 years of battery life, and 50 kbps data rate. As204
a wireless technology based on the sub-GHz frequency band,205
LoRa Technology enables massive smart IoT applications and206
solve some of the biggest challenges facing large-scale IoT,207
including energy management, natural resource reduction, pol-208
lution control, infrastructure efficiency, disaster prevention,209
and more. With an increase in the number of LoRa devices and210
networks and consequent mutual spectrum interferences, a uni-211
fied coordination-management mechanism and a large network212
are required to allocate and manage the communication spectra213
due to the unauthorized spectrum of the LoRa. Some aspects214
should be considered in LoRa applications, such as the trans-215
mission distance, the number of connected nodes, application216
scenarios, the power consumption and costs.217
C. LTE-M218
3GPP possesses three basic standards, i.e., LTE-M [13], EC-219
GSM [14] and NB-IoT [15], which are, respectively, based on220
the LTE evolution, GSM evolution and Clean Slate technolo-221
gies. Start from 3GPP R12 in 2013, LTE-Machine-to-Machine222
Fig. 2. Heterogeneous wireless-communication technologies.
(LTE-M) is intended to meet the requirements of IoT devices 223
based on the existing LTE carriers with an upstream and 224
downstream data rate of 1 Mbps [13]. LTE-M possesses four 225
basic advantages of LPWA technology including wide cov- 226
erage, massive connections, low power consumption and low 227
module costs. Due to the wide coverage, LTE-M achieves a 228
transmission gain of 15 dB in comparison to the existing tech- 229
nologies under the same licensed spectrum (700 - 900 MHz), 230
which improves the coverage ability of an LTE network sig- 231
nificantly. Besides, a LTE-M network cell can support nearly 232
100,000 connections. The standby time for LTE-M terminals 233
can be up to 10 years. 234
D. EC-GSM 235
EC-GSM (Extended Coverage-GSM) [14] was put forward 236
after the narrowband IoT technology transferring to GSM 237
(Global System for Mobile Communication), according to the 238
research project in the 3GPP GERAN (GSM EDGE Radio 239
Access Network) in 2014. As a result, a wider coverage was 240
achieved with 20 dB higher than traditional GPRS, and five 241
major objectives were proposed as follows: the improvement 242
of the indoor coverage performance, support to large-scale 243
device connectivity, simplification of the devices, reduction 244
of the power consumption and latency. However, with the 245
continuous development of the technology, the CIoT (cellu- 246
lar IoT) communication should be redefined, resulting in the 247
emergence of NB-IoT. NB-IoT is a clean-slate solution which 248
is not based on GSM. Therefore, the research of the cIoT 249
was transferred to the RAN group, that will be introduced in 250
next subsection. The EC-GSM is continued to be developed 251
by GERAN until the 3GPP R13 NB-IoT standard is almost 252
updated. 253
E. NB-IoT 254
In 2015, the 3GPP RAN initiated research on a new air- 255
interface technology called Clean Slate CIoT for narrowband 256
wireless access, which covered the NB-CIoT and NB-LTE. 257
These two technologies are fairly incompatible and compati- 258
ble with the existing LTE for easier deployment, respectively. 259
In 2016, for a unified solution, NB-IoT was considered as 260
a fusion of NB-CIoT and NB-LTE by 3GPP R13. NB-IoT 261
4 IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING
TABLE ICOMPARISON OF LPWA TECHNOLOGIES
is a new type of LPWA technology intended for sensing262
and data collection applications, such as intelligent electric263
meters, environment supervision, and etc. It can satisfy the264
requirements of non-latency-sensitive and low-bitrate appli-265
cations (time delay of uplink can be extended to more266
than 10 s, and uplink or downlink for a single user are267
supported at 160 bit/s at least), which are coverage enhance-268
ment (coverage capacity is increased 20 dB), ultralow power269
consumption (a 5-Wh batter can be used by one terminal270
for 10 years), and massive terminal access (a single sector271
can supports 50,000 connections) at transmission bandwidth272
of 200 kHz.273
F. Comparison274
We summarize the development time line of the wireless275
communication technology. Fig. 1 shows the evolution process276
of 1G, 2G, 3G, BLE, 4G, SigFox, LoRa, LTE-M, EC-GSM,277
5G, NB-IoT, etc. At the same time, we also have the joint-278
development nodes of other technologies, such as the mobile279
and short-distance communication technologies. It is clear that,280
at present, wireless-communication technology is developing281
vigorously, and represents a key node to deploy and use the282
hybrid LPWAN architecture.283
According to the IoT standards of 3GPP [16], [17] and tech-284
nology development presentation of Wei [18], we summarizes285
and compares the mentioned LPWA technologies in terms of286
the coverage, spectrum, bandwidth, data rate, and battery life287
in Table I. LTE possesses the smallest coverage because of its288
large data transmission rate and high bandwidth resources and289
energy consumption. Hence, it is not suitable for WAN appli-290
cations. LoRa does not exhibit a fixed-frequency spectrum and291
bandwidth resources, because it works in the unauthorized fre-292
quency band, and is influenced by the regulations in the area293
they are used in. As a result, it is not suitable for mobile294
communications, whereas it is suitable for wireless sensor net-295
work communications and private businesses. A low power296
consumption is the primary requirement for all LPWA tech-297
nologies, so the above technologies provide a battery life of298
up to 10 years.299
Fig. 2 compares the coverage and the data rate of300
the commonly used LPWA technology and other wireless-301
communication technologies, e.g., LoRa, EC-GSM, NB-IoT,302
WiFi, BLE (Bluetooth low-power consumption), and LTE-M.303
Short-distance and high-bandwidth communication technolo-304
gies, such as WiFi, can cover up to 100 m with a data305
transmission rate of 100 Mbps. This communication mode is306
suitable for short-distance and high-bandwidth applications. 307
For short-distance and low-data-transmission-rate communi- 308
cation technologies, e.g., Bluetooth and ZigBee, the highest 309
coverage can reach to 100 m or so while its data transfer 310
rate is 100 kbps. The communication mode is suitable for 311
short-distance, low-bandwidth applications. Long-distance and 312
high-data-transmission-rate communication technologies, such 313
as UMTS and LTE, can cover a maximum range of 10 km, 314
with a data transmission rate of 100 Mbps. This communica- 315
tion method is suitable for long-distance and high-bandwidth 316
applications. GSM can provide coverage up to 10 km, and a 317
data-transmission rate of close to 100 kbps. This communica- 318
tion mode is suitable for long-distance and medium-bandwidth 319
applications. Long-distance low-data-transmission-rate com- 320
munication technologies, such as LoRa, NB-IoT, C-IoT and 321
NB-CIoT, can cover up to 10 km, while the data transmission 322
rate is 100 kbps, and this communication method is suitable 323
for long-distance and low-bandwidth applications. 324
III. COGNITIVE LOW-POWER WIDE-AREA-NETWORK 325
ARCHITECTURE 326
Fig. 3 shows the architecture of Cognitive-LPWAN. This 327
architecture takes advantage of the LPWA communication 328
technology (in Section II), the heterogeneous IoT applications, 329
SDN and AI technology. Cognitive-LPWAN realizes the mix- 330
ing of a variety of LPWA technologies, and provides the users 331
with more efficient and convenient intelligent services. The 332
main applications of this technology are smart cities, green 333
IoT, general heterogeneous networks, as well as AI applica- 334
tions such as smart home, health monitoring, automatic driving 335
and emotional interaction. 336
(1) IoT / Heterogeneous LPWANs: It is clear from 337
the above text that, at present, there are many types of 338
wireless-communication technologies causing heterogeneous 339
and complicated IoT infrastructure and applications. As 340
shown in Fig. 3, the heterogeneous IoT platform based on 341
a variety of wireless-communication technologies, includ- 342
ing the LPWA technology. These wireless-communication 343
technologies include unlicensed LPWA technology (LoRa, 344
SigFox), short-distance wireless-communication technology 345
(BLE, WiFi), and mobile cellular-communications technology 346
(NB-IoT, LTE, 4G and 5G). The applications and coverage 347
areas supported by these technologies intersect and are widely 348
used in users’ lives. Technically, they can even replace each 349
other. However, for a comprehensive consideration of cost 350
CHEN et al.: COGNITIVE-LPWAN: TOWARDS INTELLIGENT WIRELESS SERVICES IN HYBRID LOW POWER WIDE AREA NETWORKS 5
Fig. 3. Architecture of the cognitive low-power wide-area network.
and communication performance, power consumption, mobil-351
ity [19] and other factors, the applications of these techniques352
still have their own advantages in specific scenarios. The cov-353
erages of these technologies, listed in Fig. 3, are as follows.354
The coverage of NB-IoT is < 15 km, the coverage of BLE is355
< 10 m, the coverage of WiFi is < 100 m, the coverage of356
LoRa is < 20 km, the coverage of LTE is < 11 km and the357
coverage of 5G is < 15 km [20]. The above coverage areas358
(as well as other performance indicators, such as transmission359
rate, bandwidth, sensor capacity are discussed in Section II)360
limit the popularity of these technologies throughout the field.361
Nevertheless, different applications or services have different362
requirements and use different communication technologies.363
For instance, BLE is often used for short-distance mobile-364
phone or robot communications, and WiFi is low cost, has a365
fast transmission rate and is a stable communication in WLAN366
(wireless WAN network) applications such as the smart home.367
Moreover, Nb-IoT base-stations can be built based on LTE and368
4G infrastructures, and use LTE spectrum resources, which369
greatly saves the promotion costs and supports low-power-370
consumption WAN applications. LoRa is widely used for smart371
sensing (the interconnection and transmission of mass sensors)372
and other applications thanks to its advantages in working in373
a non-authorized spectrum. The IoT and its intelligent appli-374
cations are given in details in Fig. 3. Among them, NB-IoT375
supports agriculture and environment monitoring, consumer 376
tracking, smart buildings, smart metering and smart cities. 377
BLE supports machine-to-machine (M2M) communication, 378
device-to-device (D2D) communication and other applications. 379
WiFi supports smart home, wireless access [21] and M2M 380
communication and other applications; LoRa supports smart 381
sensing, smart home, vehicle-to-vehicle (V2V) communication 382
and M2M communication. The LTE, 4G and 5G, as a wireless- 383
access technology on the Edge of the IoT, connect the above 384
IoT cells (NB-IoT, BLE, WiFi, LoRa, etc.) that are the closest 385
to the user to the Edge Cloud and realize the network-access 386
function. 387
(2) Cognitive engine: It is clear from Fig. 3 that in the 388
Edge Cloud and Cloud [22] we have introduced the cogni- 389
tive engine, deployed high-performance artificial intelligence 390
algorithms, and stored a lot of user data and IoT business 391
flows. Consequently, it can provide a high-precision cal- 392
culation and data analysis and provide Cloud support for 393
the selection of the LPWA communication technology. The 394
Cognitive engine is divided into two types, i.e., the resource- 395
cognitive engine and the data-cognitive engine [23]. 1) Data 396
cognitive engine: processing the real-time multi-modal busi- 397
ness data flow in the network environment, with data analysis 398
and business automatic processing power, executing the busi- 399
ness logic intelligently, and realizing the cognition to the 400
6 IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING
business data and resource data through a variety of cogni-401
tive calculating methods. This includes data mining, machine402
learning, deep learning and artificial intelligence. As a result,403
a dynamic guide [24] of the resource-allocation [25], [26] and404
cognitive services will be achieved. 2) The resource-cognitive405
engine can perceive the computing resources, communication406
resources [27] and network resources [28] of heterogeneous407
resources IoT [29], edge cloud and remote cloud [30], and408
make the real-time feedback of the comprehensive resources409
data to the data cognitive engine. Meanwhile, the network410
resources include the network type, data flow, communication411
quality of business and other dynamic environment parame-412
ters. Added to this, the analysis results of the data-cognition413
engine are received to guide the selection of the LPWA tech-414
nologies and real-time dynamic optimization and allocation of415
resources.416
(3) AI-enabled LPWA hybrid method: The AI-LPWA hybrid417
method is a key component of the new architecture proposed418
in this paper, as shown in the circular flow chart in Fig. 3.419
In this section we discuss the algorithm flow. When a user420
or device in the IoT makes a request, the business flow is421
transferred to the edge of the IoT through the current LPWA422
technology. It is then transmitted by the LTE, 4G, 5G [31]423
and other technologies to the Edge Cloud (wireless access424
point, router, base-station and other Edge computing nodes).425
If its request computation amount exceeds the Edge Cloud’s426
capability, it will be forwarded to the Cloud [32] again by the427
Edge computing node. In addition, the question is whether the428
Edge Cloud or the Cloud dealing with the business flow, the429
data-cognitive engine we deployed in each compute node will430
perceive all the information contained in the business flow, and431
consolidate the perceived requests, including the application432
request, content request, data volume, communication capac-433
ity, user mobility, current LPWA technology, and transmission434
rate. Subsequently, the data-cognitive engine will intelligently435
analyze the perceived requests, and extract the traffic pattern436
to transmit it to the resource-cognitive engine for the admis-437
sion control. When the Edge Cloud computing node/Cloud438
computing node admits the requests, the cognitive engine will439
allocate the computing resources for the IoT users or devices440
launching the request, and determine what type of LPWA441
wireless-communication technology will be adopted to carry442
the information back. The returning information (business con-443
tent and control information) includes the LPWA technology444
selection, interaction result, content feedback, services feed-445
back, resource allocation, and real-time monitoring. Among446
them, the selected LPWA technology will be used in the next447
request until the feedback traffic pattern of the reverse prop-448
agation in the next traffic monitoring does not conform to449
the current technological performance. At this point it will450
be replaced. The specific AI-enabled LPWA hybrid method451
modeling will be introduced in Section IV.452
IV. AI-ENABLED LPWA HYBRID METHOD MODELING453
This section presents the mathematical model of the454
AI-LPWA hybrid method. Here, we introduce the no-label455
learning algorithm for modeling [33]. Specifically, in this456
paper, aiming at the current business flow reaching the com- 457
pute nodes, its traffic patterns are extracted and added to 458
the traffic-pattern data set already there in the data-cognitive 459
engine. Next, we integrate the existing data to predict and 460
evaluate the selection of wireless-communication technologies 461
in this business. We assume that the existing traffic-pattern 462
data set (i.e., the labeled data sets) are x l = (x l1, x
l2, . . . , x
ln), 463
where x represents the traffic pattern (including the com- 464
munication time delay, data volume, transmission rate, user 465
mobility, and computing complexity), n represents the num- 466
ber of labeled data. The service request (business flow) data 467
set (i.e., unlabeled data set) that reaches the compute node is 468
xu = (xu1 , xu
2 , . . . , xum), where m represents the number of 469
non-labeled data. The label corresponding to the label data set 470
is y l = (y lx1
, y lx2
, . . . , y lxn ). Suppose the probability of assign- 471
ing some wireless communication technology to this traffic 472
flow is yuxi = {p1
xui, p2
xui, . . . , pc
xui}. Here, pj
xui
represents the 473
probability that the traffic pattern xui is predicted as category 474
j, j represents the wireless-communication technology, where 475
j = {LTE, NB−IoT, LoRa, Sigfox, BLE, WiFI, 4G, 5G. . . }, c 476
represents the number of j. Then, the pre-selection probability 477
entropy of the communication technology of the non-labeled 478
data set is shown as Eq. (1). 479
E(yuxi
)= −
c∑
j=1
pjxuilog
(pjxui
)(1) 480
It is clear from the above formula that decreasing the 481
entropy value results in a lower prediction uncertainty for the 482
new annotated data (i.e., the wireless-communication technol- 483
ogy used to allocate the business flow). Therefore, entropy can 484
be used as the pre-selection standard of the LPWA technol- 485
ogy. However, the selection of the threshold value is uncertain, 486
i.e., the low entropy value might still lead to the wrong 487
selection of the communication technology (which may be 488
directly reflected by the user experience data, such as interac- 489
tive delay and energy consumption). If the pre-selection of the 490
LPWA technology is not accurate, i.e., this traffic pattern is 491
added to the existing data set as trusted data, this can lead to 492
the accumulation of errors in the subsequent training model. 493
Specifically, the label corresponding to the traffic-pattern data 494
was incorrectly marked, which results in increased noise added 495
to the data set and increases the forward propagation error. 496
In order to overcome this problem, we apply the traffic mon- 497
itoring to the business flow, i.e., each time there is a new data 498
joining the training set, the data joining each time will be eval- 499
uated. However, we might not always trust low-threshold data. 500
Assume the independent tag data set added based on the low- 501
entropy threshold is z. Then, for any xui ⊆ z , the following 502
conditions should be met, as shown in Eq. (2). 503
E(yuxi
)� E
(y lxi
)(2) 504
In other words, the data of the error markers can be cor- 505
rected through a re-evaluation, i.e., reverse propagation in 506
Fig. 3 (4), and the tag results might be fine-tuned to reduce 507
the error. 508
CHEN et al.: COGNITIVE-LPWAN: TOWARDS INTELLIGENT WIRELESS SERVICES IN HYBRID LOW POWER WIDE AREA NETWORKS 7
Fig. 4. Comparison of the transmission delay between the AI-enabled LPWAhybrid method and different wireless-communication technologies and in theAIWAC system.
V. TESTBED AND ANALYSIS509
We use the AIWAC system [34] as an AI-enabled, heteroge-510
neous, low-power, wide-area-network architecture (application511
scene), which realizes the emotional interaction between the512
users and the smartphone / AIWAC robot. In our real experi-513
mental environment, smart clients (smartphones, robots, etc.)514
represent the IoT devices, and the local server represents the515
deployed edge computing node. Meanwhile, the cloud deploys516
a big-data center and an analytics server as the remote ser-517
vice background of the smart client. Some of the hardware518
parameters include the Smart Client (Android 7.0, 8-core,519
1.2 GHz frequency), Local Server (CentOS 7, Quad-core,520
3.4 GHz frequency), GPU Analytics Server (Ubuntu 16.04,521
NVIDIA GTX1080ti*2) and Big data Center (Android 4.4,522
16core 4 GHz + 42core 4 GHz). The computation capability523
of the Smart Client (Edge Computing Node) < Local Server524
(Fog Computing Node) < Big Data Center (Cloud).525
We deployed several wireless-communication technologies,526
which include NB-IoT, BLE, WiFi and LTE, as mentioned527
in Section IV on smartphones (communication protocol alter-528
natives) and AIWAC robots (hardware circuit-board commu-529
nication modules). These technologies represent the current530
typical LPWA communication protocols and provide us with531
a performance basis for the communication protocols in a real532
environment for the proposed Cognitive-LPWAN. In addition,533
we deployed the AI-enabled LPWA hybrid method on the cog-534
nitive engine of the Smart Client, Edge Computing Node and535
Cloud Computing Node to test the interaction performance536
comparison between the proposed scheme and the one under537
the single communication mode.538
Fig. 4 and Fig. 5 show the experimental results of539
the transmission delay and energy consumption with the540
data-transmission amount (network throughput) of different541
wireless communication technologies. Due to the narrow542
bandwidth of the NB-IoT and BLE technologies, their low543
transmission rate leads to a large interaction delay. In par-544
ticular, they are seriously defective when transferring large545
Fig. 5. Comparison of the energy consumption between the AI-enabledLPWA hybrid method and different wireless-communication technologies andin the AIWAC system.
content or files. However, they consume ultra-low energy 546
consumption while transmitting small amounts of data, so 547
they are often used in the low-power-consumption business. 548
Data-transmission rates are an advantage for WiFi and LTE 549
technologies, so transmission delays are less important than 550
the two mentioned above. However,the energy consumption 551
of LTE and WiFi are opposites. With the increase in energy 552
consumption and costs, they represent a better option for busi- 553
nesses that are sensitive to the time delay. The transmission 554
delay of the proposed AI-enabled LPWA hybrid method is 555
under AIWAC testbed and is basically consistent with WiFi 556
and LTE. This is due to the fact that the AIWAC system pro- 557
vides the emotion recognition and interactive services. And 558
it is a time-delay-sensitive application. For the loss under the 559
compromise between the energy consumption and delay of the 560
choice, it is better to use a high-speed communication tech- 561
nology. For the compromise between the energy consumption 562
and the time delay, the loss of the partial energy consumption 563
is chosen as the cost, so it is preferential to be used in high 564
speed communication technology. 565
VI. OPEN ISSUES 566
Although Cognitive-LPWAN architecture is proposed in 567
this paper, business shunt problems are discussed and the 568
AI-enabled LPWA hybrid method was introduced, in the large- 569
scale use of the hybrid LPWA technologies, the following 570
challenging problems should be solved. 571
(1) Security requirements: There are many similarities and 572
differences between the security requirements of Cognitive- 573
LPWAN and conventional IoT. The security requirements of 574
Cognitive-LPWAN mainly includes the hardware for the low- 575
power IoT, the network communication mode, and the actual 576
business requirements related to the equipment and other 577
aspects. For instance, any tiny security breach could cause 578
serious safety accidents in the IoT platform with a very large 579
number of users [35], [36]. 580
8 IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING
(2) Energy consumption and real-time management: Since581
LPWA technology has mostly been applied to the low-power582
long-distance service, its transmission rate is much lower583
than other technologies. In the heterogeneous IoT, consider-584
ing the mobility of the users and the diversity of applications,585
we should compromise the energy consumption and real-586
time performance of multiple intelligent services. Accordingly,587
we introduce short-distance communication technologies, e.g.,588
BLE and WiFi, as well as mobile cellular networks, e.g.,589
4G and 5G. We therefore meet the different requirements of590
various businesses for energy consumption and real time.591
(3) Spectrum resources optimization: Since the LPWAN592
includes technologies working in an unauthorized spectrum593
(taking LoRa and SigFox, for example), limited spectrum594
resources are required to serve a large number of IoT devices.595
Hence, there are conflicts between the bandwidth occupation596
and the spectrum resources for a large number of commu-597
nication users. Therefore, we need to use AI technologies598
such as machine learning to conduct concurrent control and599
schedule to similar users in the channel allocation, interfer-600
ence management, transmission power optimization and other601
aspects.602
(4) Infrastructure deployment: The Cognitive-LPWAN603
deployment with distributive algorithms is a hot research topic604
at present. How to build a new LPWAN architecture on the605
layout of the existing infrastructure and achieve the goal of606
minimizing cost and maximizing utilization rate is the key607
point that telecom operators need to consider.608
VII. CONCLUSION609
The diversity of the Internet of Things (IoT) in terms of610
demand and technology has led to the heterogeneity of the611
network structure and the instability of the design scheme.612
This paper has compared the performance of the commonly613
used wireless-communication technologies. In particular, we614
considered different aspects, advantages and disadvantages of615
the LoRa, SigFox, LTE-M, EC-GSM and NB-IoT technologies616
which are typical in LPWA technology. Their advantages and617
disadvantages in terms of coverage, frequency spectrum, band-618
width and data rate, and battery life are discussed. In LPWA619
and short-distance communication technologies (BLE, WiFi)620
as well as mobile cellular-network technology (4G, 5G) are621
summarized. We aimed at wireless-communication technolo-622
gies and heterogeneous networks, including smart cities and623
green IoT, as well as AI applications including smart home,624
health monitoring, automatic driving and emotional interac-625
tion. We next proposed Cognitive-LPWAN, which realizes the626
mixing of a variety of LPWA technologies and provides more627
efficient and convenient smart services. In addition, on the628
basis of the powerful cognitive computing (i.e., service aware-629
ness at the user level, intelligent transmission at the network630
level and a large data-analysis ability in the cloud) provided631
by the AI technology, we proposed the AI-enabled LPWA632
hybrid method from the perspective of traffic control. It uses633
the AI algorithms to conduct business shunt and filtering to634
wireless-communication technologies, intelligent applications635
and services for the choice of the wireless-communication636
technologies. Then, we took the AIWAC emotion interac- 637
tion system as an example and built the architecture of 638
Cognitive-LPWAN to test the proposed AI-enabled LPWA 639
hybrid method. The experimental results show that our scheme 640
can meet the demands of the communication (transmission) 641
delay (and energy consumption) in applications where appro- 642
priate communication technologies are chosen to achieve a 643
better interaction experience. Finally, we presented different 644
aspects of future research regarding the security requirements, 645
energy consumption and real-time management, spectrum 646
resource optimization and infrastructure deployment. 647
REFERENCES 648
[1] K. Hwang and M. Chen, Big Data Analytics for Cloud/IoT and Cognitive 649
Computing. Chichester, U.K.: Wiley, 2017. 650
[2] A. Asadi, V. Mancuso, and R. Gupta, “DORE: An experimental frame- 651
work to enable outband D2D relay in cellular networks,” IEEE/ACM 652
Trans. Netw., vol. 25, no. 5, pp. 2930–2943, Oct. 2017. 653
[3] R. Vaze and S. K. Iyer, “Capacity of cellular wireless network,” in Proc. 654
WIOPT, 2017, pp. 1–8, doi: 10.23919/WIOPT.2017.7959907. 655
[4] Huawei, 2016 Is the Key Year for the Development of NB- 656
IoT Industry, MWC, Barcelona, Spain, 2016. [Online]. Available: 657
http://iot.ofweek.com/2016-02/ART-132209-8120-29069354.html 658
[5] H. Lu, Y. Li, M. Chen, H. Kim, and S. Serikawa, “Brain intelligence: 659
Go beyond artificial intelligence,” Mobile Netw. Appl., vol. 23, no. 2, 660
pp. 368–375, 2017. 661
[6] F. R. Yu, “Connected vehicles for intelligent transportation systems,” 662
IEEE Trans. Veh. Technol., vol. 65, no. 6, pp. 3843–3844, Jun. 2016. 663
[7] X. Ge et al., “Wireless fractal cellular networks,” IEEE Wireless 664
Commun., vol. 23, no. 5, pp. 110–119, Oct. 2016. 665
[8] R. Usman, P. Kulkarni, and M. Sooriyabandara, “Low power wide area 666
networks: An overview,” IEEE Commun. Surveys Tuts., vol. 19, no. 2, 667
pp. 855–873, 2nd Quart., 2017. 668
[9] M. Chen, Y. Miao, Y. Hao, and K. Hwang, “Narrow band Internet of 669
Things,” IEEE Access, vol. 5, pp. 20557–20577, 2017. 670
[10] M. Saari, A. M. Baharudin, P. Sillberg, S. Hyrynsalmi, and W. Yan, 671
“LoRa—A survey of recent research trends,” in Proc. MIPRO, Opatija, 672
Croatia, May 2018, pp. 872–877. 673
[11] G. Wang, W. Xiang, and J. Yuan, “Outage performance for compute- 674
and-forward in generalized multi-way relay channels,” IEEE Commun. 675
Lett., vol.16, no. 12, pp. 2099–2102, Dec. 2012. 676
[12] B. Vejlgaard et al., “Coverage and capacity analysis of SigFox, LoRa, 677
GPRS, and NB-IoT,” in Proc. IEEE VTC Spring, Sydney, NSW, 678
Australia, Jun. 2017, pp. 1–5. 679
[13] R. Ratasuk, N. Mangalvedhe, A. Ghosh, and B. Vejlgaard, “Narrowband 680
LTE-M system for M2M communication,” in Proc. IEEE VTC Spring, 681
Vancouver, BC, Canada, May 2014, pp. 1–6. 682
[14] S. Lippuner, B. Weber, M. Salomon, M. Korb, and Q. Huang, “EC-GSM- 683
IoT network synchronization with support for large frequency offsets,” 684
in Proc. IEEE WCNC, Barcelona, Spain, Apr. 2018, pp. 1–6. 685
[15] Y. Miao, W. Li, D. Tian, M. S. Hossain, and M. F. Alhamid, 686
“Narrowband Internet of Things: Simulation and modeling,” IEEE 687
Internet Things J., vol. 5, no. 4, pp. 2304–2314, Aug. 2018. 688
[16] 3GPP. Standards for the IoT. Accessed: Dec. 2, 2016. [Online]. 689
Available: http://www.3gpp.org/news-events/3gpp-news/1805-iot_r14 690
[17] P. Reininger, “3GPP standards for the Internet-of-Things,” Huawei, 691
Shenzhen, China, Rep. 3GPP RAN WG 3, 2016. 692
[18] J. Wei, Development Status of 3GPP NB-IoT Internet of Things 693
Technology, Inst. Inf. Ind., Taipei, Taiwan, 2016. 694
[19] G. Rizzo, V. Mancuso, S. Ali, and M. A. Marsan, “Stop and forward: 695
Opportunistic local information sharing under walking mobility,” Ad Hoc 696
Netw., vol. 78, pp. 54–72, May 2018, doi: 10.1016/j.adhoc.2018.05.011. 697
[20] W. Xiang, K. Zheng, and X. Shen, 5G Mobile Communications. Cham, 698
Switzerland: Springer, 2017. 699
[21] X. Wang, M. Chen, V. C. M. Leung, Z. Han, and K. Hwang, “Integrating 700
social networks with mobile device-to-device services,” IEEE Trans. 701
Service Comput., to be published, doi: 10.1109/TSC.2018.2867437. 702
[22] Q. Zhang and S. Guo, “Online shuffling with task duplication in the 703
cloud,” ZTE Commun., vol. 15, no. 4, pp. 38–42, Oct. 2017. 704
[23] M. Chen, Y. Tian, G. Fortino, J. Zhang, and I. Humar, “Cognitive 705
Internet of Vehicles,” Comput. Commun., vol. 120, pp. 58–70, 706
May 2018. 707
CHEN et al.: COGNITIVE-LPWAN: TOWARDS INTELLIGENT WIRELESS SERVICES IN HYBRID LOW POWER WIDE AREA NETWORKS 9
[24] P. Hassanzadeh, A. M. Tulino, J. Llorca, and E. Erkip, “On coding for708
cache-aided delivery of dynamic correlated content,” IEEE J. Sel. Areas709
Commun., to be published, doi: 10.1109/JSAC.2018.2844579.710
[25] H. Shan, Y. Zhang, W. Zhuang, A. Huang, and Z. Zhang, “User behavior-711
aware scheduling based on time–frequency resource conversion,” IEEE712
Trans. Veh. Technol., vol. 66, no. 9, pp. 8429–8444, Sep. 2017.713
[26] L. Jiao, A. M. Tulino, J. Llorca, Y. Jin, and A. Sala, “Smoothed online714
resource allocation in multi-tier distributed cloud networks,” in Proc.715
IEEE IPDPSW, Chicago, IL, USA, May 2016, pp. 333–342.716
[27] H. Shan, Z. Ye, Y. Bi, and A. Huang, “Genetic algorithm based717
resource management for cognitive mesh networks with real-time and718
non-real-time services,” KSII Trans. Internet Inf. Syst., vol. 9, no. 8,719
pp. 2774–2796, Aug. 2015.720
[28] L. Chen, F. R. Yu, H. Ji, and V. C. M. Leung, “Dynamic resource721
allocation in next generation cellular networks with full-duplex self-722
backhauls,” Wireless Netw., pp. 1–12, Nov. 2016.723
[29] H. Ji, L. Xi, Z. He, and W. Ke, “ Resource allocation scheme based724
on game theory in heterogeneous networks,” J. China Univ. Posts725
Telecommun., vol. 23, no. 3, pp. 57–88, 2016.726
[30] F. R. Yu, J. Liu, Y. He, P. Si, and Y. Zhang, “Virtualization for distributed727
ledger technology (vDLT),” IEEE Access, vol. 6, pp. 25019–25028,728
2018.729
[31] X. Ge, H. Wang, R. Zi, Q. Li, and Q. Ni, “5G multimedia massive730
MIMO communications systems,” Wireless Commun. Mobile Comput.,731
vol. 16, no. 11, pp. 1377–1388, Aug. 2016.732
[32] Y. Zhi, W. Ke, and H. Ji, “Delay-aware downlink beamforming with733
discrete rate adaptation for green cloud radio access network,” J. China734
Univ. Posts Telecommun., vol. 24, no. 1, pp. 26–34, 2017.735
[33] M. Chen and V. C. M. Leung, “From cloud-based communications to736
cognition-based communications: A computing perspective,” Comput.737
Commun., vol. 128, pp. 74–79, Sep. 2018.738
[34] M. Chen, Y. Zhang, Y. Li, M. M. Hassan, and A. Alamri, “AIWAC:739
Affective interaction through wearable computing and cloud technol-740
ogy,” IEEE Wireless Commun., vol. 22, no. 1, pp. 20–27, Feb. 2015.741
[35] H. Long, W. Xiang, Y. Zhang, Y. Liu, and W. Wang, “Secrecy capacity742
enhancement with distributed precoding in multirelay wiretap sys-743
tems,” IEEE Trans. Inf. Forensics Security, vol. 8, no. 1, pp. 229–238,744
Jan. 2013.745
[36] H. Li, K. Wang, X. Liu, Y. Sun, and S. Guo, “A selective privacy-746
preserving approach for multimedia data,” IEEE Multimedia Mag.,747
vol. 24, no. 4, pp. 14–25, Oct./Dec. 2017.748
Min Chen (SM’09) has been a Full Professor with749
the School of Computer Science and Technology,750
Huazhong University of Science and Technology751
(HUST) since 2012, where he is the Director of752
Embedded and Pervasive Computing Lab. He is the753
Chair of IEEE Computer Society Special Technical754
Communities on Big Data. He was an Assistant755
Professor with the School of Computer Science and756
Engineering, Seoul National University, where he757
was a Post-Doctoral Fellow for one and half years.758
He was a Post-Doctoral Fellow with the Department759
of Electrical and Computer Engineering, University of British Columbia for760
three years. He has over 300 paper publications, including 26 ESI Highly761
Cited Papers and 9 ESI Hot Papers. He has published three books enti-762
tled Introduction to Cognitive Computing (HUST Press, 2017), Cognitive763
Computing and Deep Learning (China Machine Press, 2018), and Big Data764
Analytics for Cloud/IoT and Cognitive Computing (Wiley, 2017). He has765
over 13 800 Google Scholars Citations with an H -index of 58. His research766
focuses on cognitive computing, 5G networks, embedded computing, wear-767
able computing, big data analytics, robotics, machine learning, deep learning,768
emotion detection, IoT sensing, and mobile edge computing. He was a769
recipient of the Best Paper Award from QShine 2008, IEEE ICC 2012,770
and IEEE IWCMC 2016, the Highly Cited Researcher Award at 2018, and771
the IEEE Communications Society Fred W. Ellersick Prize in 2017. His772
top paper was cited over 1500 times. He serves as a Technical Editor or773
an Associate Editor for IEEE Network, Information Sciences, Information774
Fusion, and IEEE ACCESS. He served as a leading Guest Editor for IEEE775
Wireless Communications, IEEE Network, and the IEEE TRANSACTIONS776
ON SERVICES COMPUTING. He is a Series Editor of the IEEE Journal on777
Selected Areas in Communications. He is the Co-Chair of IEEE ICC 2012-778
Communications Theory Symposium, and IEEE ICC 2013-Wireless Networks779
Symposium. He is the General Co-Chair for IEEE CIT-2012, Tridentcom780
2014, Mobimedia 2015, and Tridentcom 2017.781
Yiming Miao received the B.Sc. degree from 782
the College of Computer Science and Technology, 783
Qinghai University, Xining, China, in 2016. She is 784
currently pursuing the Ph.D. degree with the School 785
of Computer Science and Technology, Huazhong 786
University of Science and Technology. Her research 787
interests include IoT sensing, healthcare big data, 788
and emotion-aware computing. 789
Xin Jian received the B.E. and Ph.D. degrees from 790
Chongqing University, Chongqing, China, in 2009 791
and 2014, respectively, where he is an Associate 792
Professor with the College of Communication 793
Engineering. His interests include the next gen- 794
eration mobile communication, massive machine 795
type communications, and narrow band Internet of 796
Things. 797
Xiaofei Wang (M’10–SM’18) received the M.S. and 798
Ph.D. degrees from the School of Computer Science 799
and Engineering, Seoul National University in 2008 800
and 2013, respectively. He was a Post-Doctoral 801
Research Fellow with the University of British 802
Columbia, Canada, from 2014 to 2016. He is cur- 803
rently a Professor with the School of Computer 804
Science and Technology, Tianjin University, China. 805
He has published over 80 research papers in top jour- 806
nals and conferences. His research interests include 807
cooperative edge caching and D2D traffic offloading. 808
He was a recipient of the IEEE Communications Society Fred W. Ellersick 809
Prize in 2017. 810
Iztok Humar (SM’09) received the B.Sc., M.Sc., 811
and Ph.D. degrees in telecommunications from 812
the Faculty of Electrical Engineering, University 813
of Ljubljana, Slovenia, in 2000, 2003, and 2007, 814
respectively, and the second Ph.D. degree in 815
information management from the Faculty of 816
Economics, University of Ljubljana, in 2009. 817
He is currently an Associate Professor with the 818
Faculty of Electrical Engineering, where he lec- 819
turers on Design, Modeling and Management of 820
Communication Networks on graduate and postgrad- 821
uate study. He was a Supervisor of many undergraduate students and some 822
postgraduate students. His main research topics include the energy efficiency 823
of wireless networks, cognitive computing for communications with applica- 824
tions and modeling of network loads and traffic for QoS/QoE. In 2010, he 825
was a Visiting Professor and a Researcher with the Huazhong University of 826
Science and Technology, Wuhan, China, for three months. He is a member 827
of Electrical Association of Slovenia. He served as the IEEE Communication 828
Society of Slovenia Chapter Chair for ten years. 829